MathOverflow is a question and answer site for professional mathematicians. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

In Exercise 153 of my list, it is proved that the connected components $SO_2({\mathbb R})$ and $O_2^-({\mathbb R})$ of the orthogonal group are linked as curves in the three-dimensional sphere defined by $\|M\|=1$ in the operator norm.

Question. Consider the case of $n\times n$ matrices. The real orthogonal group has two connected components $SO_n({\mathbb R})$ and $O_n^-({\mathbb R})$, each one a submanifold of dimension $n(n-1)/2$. Let $s_1(M)\ge\cdots\ge s_n(M)(\ge0)$ be the singular values of a matrix $M$. We have $s_1(M)=\|M\|$. Define the set $V_n$ by the identities $s_1(M)=\cdots=s_{n-1}(M)=1$. Is it a submanifold of dimension $n^2-n+1$ ? In $V_n$, are $SO_n({\mathbb R})$ and $O_n^-({\mathbb R})$ linked ?

Nota. The exercise in $M_2({\mathbb R})$ is included in the forthcoming second edition of my book ``Matrices'' as Exercise 21 in chapter 10.

share|cite|improve this question

The dimension of $V_n$ is at most $\binom{n}{2} +n$. For $n>2$, this is less than $2 \binom{n}{2} +1$.

It is convenient to solve a more general problem. Fix $s_1 \geq s_2 \geq \cdots \geq s_n \geq 0$. Let $M(s_1, \ldots, s_n)$ be the space of matrices with singular values $s_i$. Let $d_1$, $d_2$, ..., $d_k$ be the list of multiplicites with which the $s$'s occur, so $\sum d_i = n$. For example, if the $s$'s are $(1,1,1,\ldots, 1, r)$ then $(d_1, d_2) = (n-1, 1)$.

I claim that $\dim M(s) = 2 \binom{n}{2} - \sum \binom{d_i}{2}$. Proof: The group $O(n) \times O(n)$ acts transitively on $M(s)$, so the dimension of $M(s)$ is $\dim ( O(n) \times O(n)) = 2 \binom{n}{2}$ minus the dimensional of the stabilizer of the diagonal matrix $d(s) := \mathrm{diag}(s_1, s_2, \ldots, s_{n-1}, s_n)$. We compute the Lie algebra of the stabilizer. Let $g$ and $h$ be skew-symmetric matrices. Then, to first order, $e^g d(s) e^h = d(s)$ if and only if $g d(s) + d(s) h=0$.

For every $(i,j)$, this gives the equations $g_{ij} s_j + s_i h_{ij}$ and $s_i g_{ij} + s_j h_{ij} =0$. If $s_i \neq \pm s_j$, this forces $g_{ij} = h_{ij}=0$. If $s_i = \pm s_j$ then the sign must be $+$, as the $s_i$ are nonnegative, and we get $g_{ij} = - h_{ij}$. So the Lie algebra of the stabilizer has a basis element for every $(i,j)$ with $s_i = s_j$, and we see that the dimension of the stabilizer is $\sum \binom{d_i}{2}$.

You are interested in $\bigcup_{0 \leq r \leq 1} M(1,1,\ldots,1,r)$. Each space in the union has dimension $2\binom{n}{2} - \binom{n-1}{2} = \binom{n}{2} + n-1$, except for the $r=1$ term which is even smaller. Thus, the union has dimension at most $\binom{n}{2} +n$. I have not been able to work out whether the union is a manifold near the boundary points $r=0$ and $r=1$.

One way to salvage your question would be to look at matrices whose singular values are $(1,1-r/(n-1), 1-2r/(n-1), \ldots, 1-(n-2)r/(n-1), 1-r)$, for $r \in [0,1]$.

share|cite|improve this answer
Thank you ! I agree with your calculation and your suggestion. After all, it was important to impose $s_1=1$, but for the choice of $s_2,\ldots,s_{n-1}$, I was lazy. I change my question accordingly. – Denis Serre Oct 14 '10 at 13:36
@David. Actually, the choice of smaller $s_2,\ldots$ does not work, because the corresponding set does not contain $O_n({\mathbb R})$. Orthogonal matrices have $s_1=\cdots=s_n=1$, and are characterized by this property. I'll think to how fix the question. – Denis Serre Oct 14 '10 at 13:43
But when $r=0$, we have $1-kr/(n-1)=1$. Doesn't that fix the question? – David Speyer Oct 14 '10 at 13:48
At first glance, the question makes sense if $n=1+\binom{m+1}{2}$ (that is $n=2,4,7,11,16,...$) and $V_n$ is defined by $s_1(M)=\cdots=s_m(M)=1$. – Denis Serre Oct 14 '10 at 13:57
That also works. – David Speyer Oct 14 '10 at 14:23

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.